Repository navigation
Add regulated profile for deterministic behavior - #98
Merged
Merged
Conversation
…flags Validated deployments need a row's result to be independent of how many other requests were coalesced with it. With several compiled batch sizes the scheduler picks a different program depending on queue depth, so the same input can run through different kernels. runtime.batch_sizes: largest loads only the largest declared size per input-shape variant (the others are not read and need not exist); every dispatch then runs that one program, padded with zero rows, at the cost of full-batch compute on every dispatch. runtime.xla_flags passes XLA DebugOptions overrides to every compile. Names and value types are checked against the linked XLA's proto at startup so a typo fails loudly instead of being silently ignored, and the flags join the executable cache key (only when set, so existing cache entries keep their keys). runtime.profile: regulated bundles these as defaults (largest batch only, xla_gpu_exclude_nondeterministic_ops) while anything set explicitly wins. It deliberately does not force numerics: f32; instead it logs a bannered warning naming the matmul/conv precision actually in effect, using the TF32 probe, since only f32 is invariant across GPU generations. The executable cache used to sweep the programs of any module a worker did not read. A largest-only worker sharing a bundle directory with a normal one would have deleted the smaller sizes' programs on every start, so declared but unread modules now keep their cache records. Also corrects bundles.md, which claimed a request's batch size must equal a compiled size; coalescable models pad up to the next one. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…eport numerics=tf32 only checked that the GPU could run TF32 (compute capability >= 8.0). The startup probe that shows whether TF32 is actually used was informational, so a capable worker with NVIDIA_TF32_OVERRIDE=0 in its environment, or a stack whose kernel choice stopped using TF32, would start and serve full-f32 results to a deployment validated on TF32. Under tf32 the probe now compiles through the configured pool and fails startup unless the TF32 signature is observed (or if the probe cannot run), and NVIDIA_TF32_OVERRIDE=0 is refused before the client is created so the operator sees the actual cause. The regulated-profile startup report assumed f32 was the goal and told a tf32 deployment to switch. f32 and tf32 are now both attested modes and get a bannered info line naming the precision; only auto, which guarantees neither, keeps the bannered warning. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Sources of non-determinism:
--xla_gpu_exclude_nondeterministic_opsAdd a new "regulated" profile which handles 2 and 3 automatically, and also asserts you specify the expected precision (ie the server fails to start if you declare tf32 but its not available or f32 and vice versa)
autotuning can be made deterministic via a persistent autotuning cache